diff --git a/core/application/upgrade_checker.go b/core/application/upgrade_checker.go
index e18548a4e1b2..74fcf500267d 100644
--- a/core/application/upgrade_checker.go
+++ b/core/application/upgrade_checker.go
@@ -2,6 +2,7 @@ package application
 
 import (
 	"context"
+	"errors"
 	"sync"
 	"time"
 
@@ -222,9 +223,14 @@ func (uc *UpgradeChecker) runCheck(ctx context.Context) {
 				}, nil)
 			} else {
 				err = gallery.UpgradeBackend(ctx, uc.systemState, uc.modelLoader,
-					uc.galleries, name, nil, uc.appConfig.RequireBackendIntegrity)
+					uc.galleries, name, nil, uc.appConfig.RequireBackendIntegrity,
+					gallery.WithSkipIfBackendBusy())
 			}
-			if err != nil {
+			if errors.Is(err, gallery.ErrBackendOperationInProgress) {
+				xlog.Debug("Skipping auto-upgrade because backend is already being upgraded",
+					"backend", name)
+				continue
+			} else if err != nil {
 				xlog.Error("Failed to auto-upgrade backend",
 					"backend", name, "error", err)
 			} else {
diff --git a/core/gallery/backend_operation.go b/core/gallery/backend_operation.go
new file mode 100644
index 000000000000..4d0ceea005bd
--- /dev/null
+++ b/core/gallery/backend_operation.go
@@ -0,0 +1,106 @@
+package gallery
+
+import (
+	"context"
+	"errors"
+	"path/filepath"
+	"sync"
+)
+
+// ErrBackendOperationInProgress is returned when a non-blocking backend
+// upgrade finds another upgrade already using the same backend path.
+var ErrBackendOperationInProgress = errors.New("backend operation already in progress")
+
+// UpgradeOption customizes UpgradeBackend behavior.
+type UpgradeOption func(*upgradeOptions)
+
+type upgradeOptions struct {
+	skipIfBusy bool
+}
+
+// WithSkipIfBackendBusy makes UpgradeBackend return
+// ErrBackendOperationInProgress instead of waiting for an operation already
+// upgrading the same resolved backend path. Background auto-upgrades use this
+// so an explicit upgrade remains the owner of the in-flight work.
+func WithSkipIfBackendBusy() UpgradeOption {
+	return func(options *upgradeOptions) {
+		options.skipIfBusy = true
+	}
+}
+
+type backendOperationEntry struct {
+	token chan struct{}
+	refs  int
+}
+
+type backendOperationCoordinator struct {
+	mu      sync.Mutex
+	entries map[string]*backendOperationEntry
+}
+
+func newBackendOperationCoordinator() *backendOperationCoordinator {
+	return &backendOperationCoordinator{
+		entries: make(map[string]*backendOperationEntry),
+	}
+}
+
+var backendOperations = newBackendOperationCoordinator()
+
+// acquire serializes upgrades by their normalized backend path. refs counts
+// both the holder and waiters so an entry cannot be removed and replaced while
+// a waiter still refers to it.
+func (c *backendOperationCoordinator) acquire(ctx context.Context, path string, skipIfBusy bool) (func(), error) {
+	key := filepath.Clean(path)
+
+	c.mu.Lock()
+	entry := c.entries[key]
+	if entry == nil {
+		entry = &backendOperationEntry{token: make(chan struct{}, 1)}
+		entry.token <- struct{}{}
+		c.entries[key] = entry
+	}
+	entry.refs++
+	c.mu.Unlock()
+
+	acquired := false
+	if skipIfBusy {
+		select {
+		case <-entry.token:
+			acquired = true
+		default:
+		}
+	} else {
+		select {
+		case <-entry.token:
+			acquired = true
+		case <-ctx.Done():
+		}
+	}
+
+	if !acquired {
+		c.dropReference(key, entry)
+		if skipIfBusy {
+			return nil, ErrBackendOperationInProgress
+		}
+		return nil, ctx.Err()
+	}
+
+	var once sync.Once
+	release := func() {
+		once.Do(func() {
+			entry.token <- struct{}{}
+			c.dropReference(key, entry)
+		})
+	}
+	return release, nil
+}
+
+func (c *backendOperationCoordinator) dropReference(key string, entry *backendOperationEntry) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	entry.refs--
+	if entry.refs == 0 && c.entries[key] == entry {
+		delete(c.entries, key)
+	}
+}
diff --git a/core/gallery/backend_operation_internal_test.go b/core/gallery/backend_operation_internal_test.go
new file mode 100644
index 000000000000..00a7de41ddcf
--- /dev/null
+++ b/core/gallery/backend_operation_internal_test.go
@@ -0,0 +1,178 @@
+package gallery
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/mudler/LocalAI/pkg/system"
+)
+
+func waitForBackendOperationReferences(t *testing.T, coordinator *backendOperationCoordinator, path string, want int) {
+	t.Helper()
+	key := filepath.Clean(path)
+	deadline := time.Now().Add(time.Second)
+	for {
+		coordinator.mu.Lock()
+		entry := coordinator.entries[key]
+		got := 0
+		if entry != nil {
+			got = entry.refs
+		}
+		coordinator.mu.Unlock()
+		if got == want {
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatalf("backend operation references = %d, want %d", got, want)
+		}
+		time.Sleep(time.Millisecond)
+	}
+}
+
+func TestBackendOperationCoordinatorSerializesByPath(t *testing.T) {
+	coordinator := newBackendOperationCoordinator()
+	ctx := context.Background()
+	path := filepath.Join(t.TempDir(), "backend")
+
+	releaseFirst, err := coordinator.acquire(ctx, path, false)
+	if err != nil {
+		t.Fatalf("acquire first operation: %v", err)
+	}
+
+	// A different backend remains independent.
+	releaseOther, err := coordinator.acquire(ctx, path+"-other", true)
+	if err != nil {
+		t.Fatalf("acquire different backend: %v", err)
+	}
+	releaseOther()
+
+	// A background operation on the same backend backs off immediately.
+	if _, err := coordinator.acquire(ctx, path, true); !errors.Is(err, ErrBackendOperationInProgress) {
+		t.Fatalf("non-blocking acquire error = %v, want ErrBackendOperationInProgress", err)
+	}
+
+	acquired := make(chan error, 1)
+	go func() {
+		release, err := coordinator.acquire(ctx, path, false)
+		if err == nil {
+			release()
+		}
+		acquired <- err
+	}()
+
+	waitForBackendOperationReferences(t, coordinator, path, 2)
+	select {
+	case err := <-acquired:
+		t.Fatalf("second operation completed before release: %v", err)
+	default:
+	}
+
+	releaseFirst()
+	select {
+	case err := <-acquired:
+		if err != nil {
+			t.Fatalf("second operation failed after release: %v", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("second operation did not acquire after release")
+	}
+
+	coordinator.mu.Lock()
+	defer coordinator.mu.Unlock()
+	if len(coordinator.entries) != 0 {
+		t.Fatalf("coordinator retained %d idle entries", len(coordinator.entries))
+	}
+}
+
+func TestBackendOperationCoordinatorDropsCanceledWaiter(t *testing.T) {
+	coordinator := newBackendOperationCoordinator()
+	path := filepath.Join(t.TempDir(), "backend")
+	release, err := coordinator.acquire(context.Background(), path, false)
+	if err != nil {
+		t.Fatalf("acquire first operation: %v", err)
+	}
+
+	canceled, cancel := context.WithCancel(context.Background())
+	cancel()
+	if _, err := coordinator.acquire(canceled, path, false); !errors.Is(err, context.Canceled) {
+		t.Fatalf("canceled acquire error = %v, want context.Canceled", err)
+	}
+	release()
+
+	// A canceled waiter must not leave the path permanently busy.
+	reacquired, err := coordinator.acquire(context.Background(), path, true)
+	if err != nil {
+		t.Fatalf("reacquire after cancellation: %v", err)
+	}
+	reacquired()
+}
+
+func TestUpgradeBackendUsesResolvedConcretePath(t *testing.T) {
+	backendsPath := t.TempDir()
+	state, err := system.GetSystemState(system.WithBackendPath(backendsPath))
+	if err != nil {
+		t.Fatalf("get system state: %v", err)
+	}
+
+	concreteName := "concrete-backend"
+	concretePath := filepath.Join(backendsPath, concreteName)
+	if err := os.MkdirAll(concretePath, 0750); err != nil {
+		t.Fatalf("create concrete backend: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(concretePath, runFile), []byte("#!/bin/sh\n"), 0755); err != nil {
+		t.Fatalf("write concrete run file: %v", err)
+	}
+	if err := writeBackendMetadata(concretePath, &BackendMetadata{Name: concreteName, Version: "1"}); err != nil {
+		t.Fatalf("write concrete metadata: %v", err)
+	}
+
+	metaName := "meta-backend"
+	metaPath := filepath.Join(backendsPath, metaName)
+	if err := os.MkdirAll(metaPath, 0750); err != nil {
+		t.Fatalf("create meta backend: %v", err)
+	}
+	if err := writeBackendMetadata(metaPath, &BackendMetadata{Name: metaName, MetaBackendFor: concreteName}); err != nil {
+		t.Fatalf("write meta metadata: %v", err)
+	}
+
+	release, err := backendOperations.acquire(context.Background(), concretePath, false)
+	if err != nil {
+		t.Fatalf("hold concrete backend operation: %v", err)
+	}
+	t.Cleanup(release)
+
+	err = UpgradeBackend(
+		context.Background(), state, nil, nil, metaName, nil, false,
+		WithSkipIfBackendBusy(),
+	)
+	if !errors.Is(err, ErrBackendOperationInProgress) {
+		release()
+		t.Fatalf("auto-upgrade error = %v, want ErrBackendOperationInProgress", err)
+	}
+
+	waitResult := make(chan error, 1)
+	go func() {
+		waitResult <- UpgradeBackend(context.Background(), state, nil, nil, metaName, nil, false)
+	}()
+	waitForBackendOperationReferences(t, backendOperations, concretePath, 2)
+	select {
+	case err := <-waitResult:
+		release()
+		t.Fatalf("manual upgrade returned while concrete backend was busy: %v", err)
+	default:
+	}
+
+	release()
+	select {
+	case err := <-waitResult:
+		if err == nil || errors.Is(err, ErrBackendOperationInProgress) {
+			t.Fatalf("manual upgrade error after release = %v, want normal gallery lookup failure", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("manual upgrade did not continue after concrete backend was released")
+	}
+}
diff --git a/core/gallery/upgrade.go b/core/gallery/upgrade.go
index 7abc65a4358b..ed9f5f341d87 100644
--- a/core/gallery/upgrade.go
+++ b/core/gallery/upgrade.go
@@ -240,7 +240,7 @@ func summarizeNodeDrift(nodes []NodeBackendRef) (majority struct{ version, diges
 
 // UpgradeBackend upgrades a single backend to the latest gallery version using
 // an atomic swap with backup-based rollback on failure.
-func UpgradeBackend(ctx context.Context, systemState *system.SystemState, modelLoader *model.ModelLoader, galleries []config.Gallery, backendName string, downloadStatus func(string, string, string, float64), requireIntegrity bool) error {
+func UpgradeBackend(ctx context.Context, systemState *system.SystemState, modelLoader *model.ModelLoader, galleries []config.Gallery, backendName string, downloadStatus func(string, string, string, float64), requireIntegrity bool, opts ...UpgradeOption) error {
 	// Look up the installed backend
 	installedBackends, err := ListSystemBackends(systemState)
 	if err != nil {
@@ -259,7 +259,32 @@ func UpgradeBackend(ctx context.Context, systemState *system.SystemState, modelL
 	// If this is a meta backend, recursively upgrade the concrete backend it points to
 	if installed.Metadata != nil && installed.Metadata.MetaBackendFor != "" {
 		xlog.Info("Meta backend detected, upgrading concrete backend", "meta", backendName, "concrete", installed.Metadata.MetaBackendFor)
-		return UpgradeBackend(ctx, systemState, modelLoader, galleries, installed.Metadata.MetaBackendFor, downloadStatus, requireIntegrity)
+		return UpgradeBackend(ctx, systemState, modelLoader, galleries, installed.Metadata.MetaBackendFor, downloadStatus, requireIntegrity, opts...)
+	}
+
+	options := upgradeOptions{}
+	for _, opt := range opts {
+		if opt != nil {
+			opt(&options)
+		}
+	}
+
+	backendPath := filepath.Join(systemState.Backend.BackendsPath, backendName)
+	release, err := backendOperations.acquire(ctx, backendPath, options.skipIfBusy)
+	if err != nil {
+		return fmt.Errorf("backend %q: %w", backendName, err)
+	}
+	defer release()
+
+	// Refresh installed state after waiting. Another upgrade may have replaced
+	// metadata while this call was blocked on the same backend path.
+	installedBackends, err = ListSystemBackends(systemState)
+	if err != nil {
+		return fmt.Errorf("failed to refresh installed backends: %w", err)
+	}
+	installed, ok = installedBackends.Get(backendName)
+	if !ok {
+		return fmt.Errorf("backend %q: %w", backendName, ErrBackendNotFound)
 	}
 
 	// Find the gallery entry. Unfiltered for the same reason as the check
@@ -285,7 +310,6 @@ func UpgradeBackend(ctx context.Context, systemState *system.SystemState, modelL
 		return fmt.Errorf("upgrade %q: %w", backendName, err)
 	}
 
-	backendPath := filepath.Join(systemState.Backend.BackendsPath, backendName)
 	tmpPath := backendPath + ".upgrade-tmp"
 	backupPath := backendPath + ".backup"
 
diff --git a/core/services/galleryop/managers_local.go b/core/services/galleryop/managers_local.go
index 70fc018e000b..b798b049886f 100644
--- a/core/services/galleryop/managers_local.go
+++ b/core/services/galleryop/managers_local.go
@@ -105,12 +105,20 @@ func (b *LocalBackendManager) ListBackends() (gallery.SystemBackends, error) {
 	return gallery.ListSystemBackends(b.systemState)
 }
 
-// UpgradeBackend ignores op.ID and op.TargetNodeID: a single-node install
-// reports progress through the local progressCb already, and there is only
-// one node to target. Both fields only matter for distributed per-node
-// streaming/scoping (see DistributedBackendManager.UpgradeBackend).
+// UpgradeBackend uses an empty op.ID to identify the UpgradeChecker's
+// background operation, which backs off instead of waiting when the same
+// backend is busy. A single-node upgrade reports progress through the local
+// progressCb already, and TargetNodeID only matters for distributed per-node
+// scoping.
 func (b *LocalBackendManager) UpgradeBackend(ctx context.Context, op *ManagementOp[gallery.GalleryBackend, any], progressCb ProgressCallback) error {
-	return gallery.UpgradeBackend(ctx, b.systemState, b.modelLoader, b.backendGalleries, op.GalleryElementName, progressCb, b.requireBackendIntegrity)
+	var opts []gallery.UpgradeOption
+	if op.ID == "" {
+		opts = append(opts, gallery.WithSkipIfBackendBusy())
+	}
+	return gallery.UpgradeBackend(
+		ctx, b.systemState, b.modelLoader, b.backendGalleries,
+		op.GalleryElementName, progressCb, b.requireBackendIntegrity, opts...,
+	)
 }
 
 func (b *LocalBackendManager) CheckUpgrades(ctx context.Context) (map[string]gallery.UpgradeInfo, error) {